基于元学习的图卷积网络少样本学习模型

Few-Shot Learning on Graph Convolutional Network Based on Meta learning

少样本学习是目前机器学习研究领域的热点和难点.针对现有的少样本学习模型不能有效捕捉数据特征与数据标签之间的联系,造成分类模型泛化能力弱的问题,提出一种基于元学习的原型空间图卷积网络少样本学习模型FSL-GCNPS(Few-Shot Learning of Graph Convolutional Network on Prototype Space).首先,利用卷积神经网络提取多任务数据的特征向量;其次,为了将特征向量映射到原型空间中,根据元学习的训练策略得到特征向量的类原型表达;然后,通过类原型向量和类向量之间的嵌入表示,构建图结构数据,并进行图卷积网络训练、推理.实验结果表明,相较于经典少样本学习方法,FSL-GCNPS模型拥有更好的分类准确率和分类稳定性.同时,在医学图像领域数据集上实验表明,FSL-GCNPS具有很好的跨域适应性.

Few shot learning is a hot and difficult problem in the field of machine learning.The existing few-shot learning model cannot effectively capture the relationships between data feature information and data label,thus causing the generalization ability of the resulting classifier would be weaker.A few-shot learning of graph convolutional network on prototype space,termed FSL-GCNPS,is developed.Firstly,the feature vectors are extracted on multi-task data by convolu⁃tional network.Secondly,in order to map the feature vectors into the prototype space,representation learning for the classes based on prototype network is proposed.Next,the graph is structured by combing the classes prototype vectors with class vectors.Then,FSL-GCNPS is trained using Meta learning.The experimental results show that FSL-GCNPS has better cross-domain adaptability in the medical image domain compared with the traditional deep learning models.Meanwhile,the FSL-GCNPS model has better classification accuracy and classification stability compared with the classical Few-shot learn⁃ing algorithm.